abstract - association for computational linguisticsmiller (1984) called 'informavores',...

Thrane, T. 1988. Symbolic Representation and Natural Language.

- 1 17 -

ABSTRACT

The notion of symbolizability is taken as the second requisite of

computation (the first being 'algorithmizability'), and it is

shown that symbols, qua symbols, are not symbolizable. This has

farreaching consequences for the computational study of language

and for Al-research in language understanding. The representation

hypothesis is formulated, and its various assumptions and goals

are examined. A research strategy for the computational study of

natural language understanding is outlined.

Symbolic Representation and Natural LanguageTorben ThraneProceedings of NODALIDA 1987, pages 117-154

- 118 -

SYMBOLIC REPRESENTATION AND NATURAL LANGUAGE

b y

Torben Thrane

Center for Informatics, University of Copenhagen

1. Introduction: algorithms and symbolic representation

The algorithm is without doubt th e fundamental concept in com

puter science. Problems that can be solved by computer are all

algorithmic: they can be presented on a form which invites a di

vision of the overall problem into constituent parts, each of

which can be solved sequentially and deterministically. When the

last constituent problem is solved the overall problem is solved.

To get a computer to solve a problem, however, its constituent

parts must be s y m h o llz a b le : they must be put on a form which is

accessible to the computer. This precondition is summed up under

the rubric 'representation'.

The inescababillty of these two preconditions is never in doubt.

However, doubts have recently been raised with respect to the

value of the comparison that is often made, explicitly or implic

itly, between the problem solving capacities of men and machines,

and in particular with respect to the role allegedly played by

representation as the constitutive feature of those capacities.

118Proceedings of NODALIDA 1987

Understanding natural language is among the problems whose solu

tion has been expected to be accessib le through computer simula

tion , p rec ise ly on the assumption o f a common representational

basis fo r the problem solving cap acities o f men and machines.

- 119 -

2. The representation hypothesis

The topic of the present paper, thus loosely outlined. Is sym b ol

i c r e p r e s e n t a t io n , in general as well as more specifically with

respect to the role it plays in computational linguistics. Init

ially, the notion of representation can be presented as a simple

formula:

(1) a R b, which reads: 'a represents b'.

Representation is the name of a relation holding between two ent

ities. The logical properties of the relation are usually taken

to be i r r e f l e x i v i t y , i n t r a n s i t i v i t y , and a sym m etry. Apart from

these logical ones, R has properties sometimes summed up by say

ing that a sta n d s f o r , c o m p lie s w ith , r e f e r s t o , s y m b o l i s e s , d e

n o t e s , d e p i c t s , d e s ig n a te s , c o r r e sp o n d s w ith b.

The logical properties of the entities between which representa

tion holds are more difficult to characterize briefly. Let us as

sume, initially, that a is a physical entity, whereas b is typol-

ogically unspecified. We return to this issue below.


In relation to (1) we can formulate an overall hypothesis, the

r e p r e s e n t a t io n h y p o t h e s is , which says:

(2) All intelligent behaviour presupposes the formula (1).

Ultimately, this hypothesis aims to explain how such systems as

Miller (1984) called 'informavores', can function as autonomous

entities in larger physical environments, which they both affect

and are affected by.

- 120 -

2 . 1 . The adequacy o f th e form ula

The formulation (1) invites the view that representation is a

contextfree phenomenon, and that it eludes situationally condi

tioned interpretation. This is incorrect. Already C.S. Peirce -

who in this connection can be considered one of the founding

fathers of representation theory - insisted on the decisive in

fluence that situation and context has on the interpretation of a

sign. And perhaps even more importantly, he insisted that even

the recognition of a physical entity as a sign presupposed back

ground, interpretation, and what he called 'semiosis' or - as we

shall say - 'the semiotic process': the process by which there is

created in an observer a mental correlate - Peirce's 'interpret-

ant' - of a physical phenomenon which, in virtue of this process,

now becomes a sign o f its object, b, to the observer.(CP 2.227-9;

Hookway, 1985:118-144). From this perspective, nothing is a sign

'in itself. A sign is c r e a te d - through the semiotic process. So

the formula (1) can be amended to:


(3) a R b for observer 0 in situation S in virtue of the con

ventions C.

This means that an internal representation of b has been created

in 0, 'corresponding to' the physical sign, a. This internal rep

resentation, according to Peirce, is itself a sign, for which a

new interpretant can be created by a recursive application of the

semiotic process, and so on, ad infinitum. By way of continuation

of the discussion of the logical properties of the entities be

tween which R holds, we get a glimpse here of a systematic vacil

lation in the conception of a in the formula: a can either be re

garded as the p h y s ic a l s ig n which represents b; or else a can be

understood as the co n cep tu a l s t r u c t u r e which has been created in

0 by virtue of O's taking a as a sign for b.

If a in this way can be thought of as either a physical or a men

tal phenomenon, b must be so conceived as well. It causes no

trouble to entertain the idea that physical entities can be rep

resented. Nor does it cause trouble to entertain the idea that

mental or abstract phenomena can be represented. There would

seem, therefore, to be no trouble in accepting that meaning can

be represented, no matter whether meaning is considered to be a

mental phenomenon or not; cf. Searle 1983:Ch.8 for a general dis

cussion of this point.

- 121 -

However, there does crop up a problem for computational linguis

tics. On the view set out above, natural language is itself a

representational system of interpreted symbols. At the same time,

natural language is - for computational linguistists - a phenom

enon that must itself be representable by computationally inter


pretable symbols. Here the intransitivity of the general repres

entation relation becomes apparent. If it were transitive it

would be child's play to construe natural language interfaces,

since then, say, a string variable, a, would appear to represent

whatever the content of a represents. But this is not how things

work. If a in this instance represents anything apart from the

sequence of alphanumerical characters that make up the string,

then it is a location in the computer's memory, and not, for ex

ample, the person that a string 'Tom Jones' is supposed to rep

resent in a given context. From one perspective, then, computa

tional representation of natural language is a special case of

hypostasis.

Emerging from this discussion is the following startling fact:

symbolic representation of an object, b, which i t s e l f has been

interpreted as a symbol, is impossible! It can be a physical

entity which - in other circumstances - c o u ld function as a sym

bol. Or, it can be a mental phenomenon, an abstract object, a

conceptual structure, in addition to whatever meaning is supposed

to be.

This conclusion can be summed up in slogan fashion:

(4) S ym bols are n o t sy m b o liz a b le

- 122 -

This slogan may have consequences for the proper study of various

linguistic phenomena, anaphora for example. Of more immediate

concern, however, is the fact that it can be construed as the

basis of the blpartltlon which characterizes previous attempts to

Justify the representation hypothesis.


2.2. S ta g es a lon g th e pa th o f th e r e p r e s e n t a t io n h y p o th e s is

Attempts to justify the representation hypothesis come from many

sides and from many more or less related academic disciplines.

The same can be said of the attempts to dismiss it as untenable,

in particular and most recently by Winograd & Flores (1986).

Let us first dismiss one obvious possibility of discrediting the

representation hypothesis. It criticizes any attempt to justify

it that takes the form of clarifying (1), correctly claiming that

justification of it should take the form of a clarification of

(3). Clarification of (1) enters into the ultimate clarification

of (3), but they are not the same; nor can clarification of (1)

ever on its own count as justification of (2). I shall therefore

assume, despite some evidence to the contrary, that everybody who

has been seeking justification of (2) in the pursuit of clarifi

cation of (1) in fact consciously, if tacitly, have been pursuing

clarification of (3).

Clarification of (3) could be regarded as an algorithmic problem

for the ultimate solution of which a number of constituent prob

lems have been formulated and described by various academic dis

ciplines and scholarly persuasions.

- 123 -

These constituent problems can with some idealization be present

ed in a hierarchical structure of mutually explanatory and 'sup

port ' systems, as shown in (5):


- 124 -

(5 ) The Representation Hypothesis explains intelligent behaviour, and is justified by -

The thesis of symbolization and symbolizability, which explains

The concept of notation system.

Reasoning as symbol- manipulation, which explains -

Brentano's problem. The acquisition and storing of knowledge, which explains -

Certain types of speech-acts.

The correspondence between language and reality, which explains -

The universe of discourse.

The correlation between expression and content, which explains -

The 'nature' of symbols.


The right-hand side comprises theories, hypotheses and presuppos

itions which form essential components of the overall Representa

tion hypothesis, whereas the left-hand side displays (examples

of) concrete problems or questions, the answers to which presup

pose the validity of the thesis in question. In what follows I

will discuss the items on the right-hand side. The items on the

left-hand side will be considered in the form of a series of di

gressions, which together will outline a research strategy for

man-machine interaction in natural language.

- 125 -

3. The constituent parts of the representation hypothesis

The list of constituent problems under (5) reads like a catalogue

of a significant portion of Western philosophy, cognitive psych

ology and epistemology in terms of substance. We shall be con

cerned with them only from the computational point of view.

3.1. P h y s ic a l sym bol sy s te m s

The bipartition mentioned above, and the vacillation in the con

ception of a mentioned in 2.1., has led to a bifurcation of com

putational research which both proceed from Newell's (1980; New

ell & Simon 1976) account of a physical symbol system.

A physical symbol system is a system which subscribes to the laws

of physics and which at any given time contains a set of struc-


tured objects called 'expressions' or 'symbol structures'. Each

expression is constituted by a number of sy m b o ls , ordered in a

principled way. Symbols are physical objects whose internal

structure may be quite complex, but which have the common prop

erty of being combinable with other symbols to form expressions.

In addition, a physical symbol system comprises a set of process

es that may create, change, destroy, and reproduce expressions. A

physical symbol system, then, is a machine which produces a con

tinuous, but continually changing, stream of symbol structures.

The precondition for the proper functioning of a physical symbol

system is the notion of i n t e r p r e t a t i o n , as defined in computer

science (Newell 1980:158). Interpretation is there understood as

the act of accepting as input an expression which represents a

process, and then executing that process.

- 126 -

The first of the two directions of computational research alluded

to above studies the internal structure of symbols with a view to

establishing a typology of symbols of greater perspicuity and

more practical versatility than Peirce's, for example in connec

tion with the generation of fonts and character sets (cf. Knuth

1982; Hofstadter 1982), the creation of MM interfaces, document

representation (cf. Levy, Brotsky & Olson 1987; Southall 1987),

the development of graphics systems, etc. This direction, under

the label 'figural representation', is deeply influenced by the

art-aesthetic discussion of the concept of representation, which

rejects s i m i l a r i t y between a and b as the proper grounds for rep

resentation in favour of substitution or functional equivalence.

Consider in this connection Picasso's famous reply to contempor

ary criticism that his portrait, Gertrude, did not resemble Mrs.

Stein: "Don't worry. It will".


The other direction seeks to explain the behaviour of informa-

vores. This direction proceeds from cognitive psychology as much

as from semiotics - it is known as 'cognitive science' which sub

sumes the more practical study of artificial intelligence. This

was Newell's own main interest. He formulates the following hy

pothesis :

(6) The P h y s ic a l Symbol S ystem H y p o th e s is

The necessary and sufficient conditions for a physical system

to exhibit general intelligent action is that it be a physic

al symbol system.

Newell, 1980:170

'Necessary' and 'sufficient' in this connection mean, respective

ly, that a system displaying what we would be prepared to call

intelligent behaviour, will always upon closer inspection turn

out to be a physical symbol system; and a physical symbol system

of sufficient size will always be amenable to organization in

such a way that it will display behaviour that we would be pre

pared to call intelligent.

- 127 -

Clearly, interest in the properties of the symbol differs accord

ing to which of the two directions one follows. In the first in

stance we get an interest in the physical properties of symbols

that may alleviate their e x te n s io n a l interpretation. In the sec

ond, interest centres around the physical properties of symbols

that will secure a particular in te n s io n a l interpretation. In

these terms the semiotic process can be seen as a complex series


of steps whereby a particular physical object undergoes various

internal processes (Newell calls them 'symbolic')/ whereby they

are turned into meaningful structures, le structures that det

ermine the symbol system's subsequent behaviour. It is important

to realize that we have no immediate access to these structures,

but only recognize them through the behaviour of the system.

Consequently, we can only try to describe them on the basis of an

abstraction from observed behaviour, in an appropriate formal no

tation, if the need arises. In this way, both directions take a

crucial interest in the physical properties of symbols which

converges in the need for interpretation, extensional and

intensional. This leads to a digression on notational systems.

- 128 -

3.1.2. D i g r e s s i o n : N o ta tio n a l sy s te m s

The fundamental property of a symbol is that it is an object man

ifest to the senses. But - as Newell made clear - a symbol may be

of a complex internal structure. This structure will in some cas

es be amenable to description by means of a notational system,

v i z those cases where the atomic parts of every symbol in the

scheme constitute a set that satisfies the five requirements on

notational systems formulated by Nelson Goodman (1976):

(7)(a) S y n t a c t ic d i s c r e t e n e s s : the decision whether some arbitra

ry inscription belongs to a particular character and not

another is deterministic;

(b) S y n t a c t ic d i s j o i n t n e s s : any inscription either belongs to

one and only one character, or does not belong to the

scheme;


(c) U nam higuity: any character, as well as any inscription of

any character, must be unambiguous;

(d) Sem antic d i f f e r e n t i a t i o n : the decision whether some ref

erent of a particular inscription of any character belongs

to one class of objects or another is deterministic;

(e) Sem antic d i s j o i n t n e s s : no two characters in a notational

system can have any referent in common.

- 129 -

In (8) are displayed samples of signs that are amenable to in

ternal description on the basis of a notatlonal system (a), and

of signs the internal structure of which does not reflect a no-

tational system, but which must rather be described on the basis

of iconicity (b):


- 130 -

(8)(a)

/* Insert page with figures (8)(a) and (b) and remove this line*/

(b)


The interesting thing about these claims in our connection is

that the alphabet satisfies them, whereas larger linguistic enti

ties as a rule do not. The I n te r n a tio n a l P h o n e tic s A s s o c ia t io n

notation in fact subsumes both types: a subset - used for phonem

ic transcriptions - forms a notational system on the criteria

above, the scheme as a whole - used for phonetic description -

does not. Semantic networks, frames, etc., do not constitute not

ational systems in the required sense, the propositional and

predicate calculi do. And finally, all (procedural?) programming

languages are notational systems. The parenthesis indicates some

hesitation with respect to programming languages like Prolog and

Lisp, and many tools specifically developed as system-building

aids for knowledge engineering. And the hesitation is due to un

certainty whether to regard such languages from the point of view

of the programmer or from the point of view of the machine in

making the decision. This ties in with the representation hier

archy (see below, 3.2.1).

- 131 -

3.2. R ea son in g as sytn bolm anipulation

..Reason, when wee reck o n i t amongst

th e F a c u l t i e s o f th e mind ... is no

th in g b u t Reckoning ( th a t i s Adding

and S u b tr a c tin g ) o f th e C onsequences

o f g e n e r a ll names a greed upon, f o r

th e marking and signifying o f our

th o u g h ts .


This passage, from Thomas Hobbes L evia th a n (1651:1.5), is one of

the earliest expressions of what Winograd & Flores somewhat

sweepingly style the 'rationalistic tradition' in representation

theory. In our own time, Johnson-Laird (1983:2-4) credits Kenneth

Craik with the first full-fledged modern version. He describes

reasoning as a process that falls into three phases:

(9)(a) A 'translation' of some external process into an internal

representation in terms of words, numbers or other sym

bols;

(b) The derivation of other symbols from them by some sort of

inferential process;

(c) A 'retranslation' of these symbols into actions, or at

least a recognition of the correspondence between these

symbols and external events, as in realizing that a pre

diction is fulfilled.

Johnson-Laird (1983:2-3)

The symbol has become a mental code with psychological reality, a

view which harks back to Peirce's notion of 'interpretant'. How

ever, we should be wary of identifying the two views, mainly be

cause cognitive scientist are more interested than Peirce in ex

plaining the b e h a v io u r of a system on the basis of representa

tional (semantic) content; cf. Pylyshyn's (1984:39) formulation:

- 132 -

In cognitive science,..., we want something stronger than

derived semantics, inasmuch as we want to e x p la in th e s y s

te m 's b e h a v io r b y a p p ea lin g to th e c o n te n t o f i t s r e p r e s

e n t a t i o n s .

(my Italics)


The justification of this thesis is central to cognitive psychol

ogy, for if it can be justified, Brentano's problem disappears.

- 133 -

3 . 2 . 1 . D ig r e s s io n : B r e n ta n o 's problem

How can physical stimuli determine the behaviour of a biological

system even though no direct causal relationship exists between

stimulus and behaviour? This is a nutshell formulation of the

classical problem which Brentano attempted to solve by introduc

ing a distinction between an 'object' and our mental 'represent

ation' of it, and which, since then, has been one of the consti

tutive problems of cognitive psychology, on a par with philos

ophy's mind-body problem.

Rather than solve it, cognitive science believes to have d i s

s o lv e d it - with reference to the physical embodiment of symbol

systems and the representation hierarchy (Newell 1980:172-75,

1982; Haugeland 1982; Johnson-Laird 1983:399ff; Pylyshyn 1984;

but cf. also Winograd & Flores 1986:86-89 and Ch. 8).

A computer can be exhaustively described in various ways: as a

physical device, as a collection of logically interpreted cir

cuits, as a deterministic sequence of states defined by a pro

gram, etc. The notion of a hierarchy of representational levels

stems from the realization that, depending on which type of de

scription is chosen, there is a corresponding, radical change in

the proper description of the nature of the information-proces

sing relevant to that type: as electrical current switching on


and o f f , as interpretation of particular electrical patterns as

logical values, as interpretation of larger chunks of such pat

terns as alphabetic characters or numbers, of yet larger chunks

as lists of objects or strings of characters, etc.

There is no disagreement on these principles in so far as they

are used to describe what com p u ters do. But if the account is

used as a metaphor - or, indeed, as a literal description - of

what people do in processing information from physical stimuli of

the senses, disagreement is fierce. Searle (1980) draws a useful

distinction between 'weak' and 'strong' artificial intelligence

research in a discussion of these matters. Weak AI is character

ized by regarding computers and programs as tools that enable us

to test hypotheses about cognitive processes in a rigorous man

ner, whereas - in strong AI - the appropriately programmed com

puter provides the explanation of such processes, with a signi

ficant parallel being posited between the Jump from hardware to

software (in the case of the computer) and from brain to mind (in

the case of people). But - to the strong AI researcher (eg Pylys-

hyn) - the point is precisely that it is impossible to define a

definite line across which this jump is made, as suggested by the

fluid, yet perfectly specifiable, levels of the representation

hierarchy.

- 134 -

The representation hypothesis, in conjunction with the strong AI

interpretation of the representation hierarchy, provides a suffi

ciently rich explanation of our interaction with our environment

to merit serious consideration. Attempts to discredit it, there

fore, must provide a plausible, and equally rich, alternative

solution to Brentano's problem to be creditable themselves. This


is what Winograd 6 Flores (1986:Ch. 4) do in their appeal to the

cognitive theories of the Chilean biologist Maturana.

- 135 -

3 . 3 . The a c q u is it io n and s to r in g o f knowledge

The problem of how we a cqu ire knowledge is at the epistemological

core of Western philosophy. The problem of how we store it, once

acquired, has only become of prime importance with the advent of

computers, for if there is one thing on which all AI researchers

agree it is that knowledge is extremely bulky. So, although the

question of how best to feed the necessary information into the

system is of concern to AI research in general, the burning ques

tions are rather how to cope with its bulk without loss, and how

to preserve it in a form suitable for rapid access and retrieval.

These problems crucially involve matters of representation, and

specific knowledge r e p r e s e n ta tio n languages have been developed

to cope with them; cf. Waterman (1986:339-365) for a survey.

All attempts to create knowledge representation languages have

assumed the validity of the knowledge r e p r e s e n ta tio n h y p o th sis ,

first explicitly formulated by Smith (1982). It goes like this:

(10) Any mechanically embodied intelligent process will be com

prised of structural in g r e d ie n ts that a) we as ex te rn a l ob

servers n a tu r a lly take to r e p r e se n t a p r o p o s it io n a l account

o f th e knowledge that the o v e r a ll p r o c e s s e x h i b i t s , and b)

independent of such external semantic attribution, play a

form al but causal and essential role in engendering the b e

haviour th a t m a n ifests th at know ledge.

Qu.f. Brachman & Levesque (1985:33

(mv italics)


This formulation comprises a series of quite central claims about

the nature, organization, and function of the knowledge required

for the performance of rational behaviour.

Firstly, knowledge is r e p r e s e n ta b le - or symbolizable - by s t r u c

tu r a l elements. Accessible knowledge is assumed to be organized

along previously determined patterns or principles. Only access

ible knowledge determines behaviour.

Secondly, the representation of knowledge structured along these

lines is assumed to consist of a collection of p r o p o s i t i o n s ,

which we can define here as truth-valued abstractions over states

of affairs.

Thirdly, it is supposed to be n a tu ra l for us to see the situation

in this light.

Fourthly - and in direct continuation of the digression on Bren-

tano's problem above - it is the knowledge, structured in this

way, that is the cause of the system's behaviour, and which we

call intelligent because it reflects a reasonable 'awareness' of

the accessible knowledge. Behaviour is the only external (or in

terpersonal) evidence of accessible knowledge.

- 136 -

Finally, the triggering of rational behaviour is strictly formal,

which means that it is not the propositional c o n te n t as such

which is the factor determining behaviour, but rather the struc

tural occurrence of a particular configuration of truth-values in

the overall knowledge structure. Pylyshyn's (1984) major aim is

to escape this conclusion.


No doubt all of these claims - and their consequences - merit

discussion, but I will stick to just one, viz. that acquired

knowledge is structured propositionally, and I will do so by way

of yet another digression, on speech acts.

- 137 -

3.3.1. D ig r e s s io n : Speech A c ts

To make a statement, to ask a question, to issue an order are the

three major types of speech act, in the sense that most languages

make distinctions in their grammatical systems between the types

of sentences typically used to perform them: declarative, inter

rogative, and imperative, respectively. Among these, declarative

sentences have had a particularly prominent position in theoret

ical discussions of semantics, because they are natural language

expressions of streamlined, truth-valued propositions, and be

cause theoretical semantics has often seen it as its major busi

ness to account for the conditions that make a particular declar

ative sentence true.

If, however, the major semantic business is to account for the

circumstances in which a particular sentence can be said to have

been u n d e rs to o d , priorities change. Documentation of understand

ing is a c t i o n : documentation of the understanding of a question

is a suitable locutionary act, documentation of an order is exe

cution of a suitable physical act, locutionary or not. These are

fairly clear. But what action do we perform in order to document

understanding of a declarative sentence?


One possible answer supports the claim above, that knowledge is

stored and structured on propositional form. Informally, it says

that anyone who hears a declarative sentence will carry through a

recursive check as to whether the proposition expressed by it is

already part of his 'knowledge base' or not. If it is, and if

there is no new evidence for altering one's knowledge on this

point, the sentence is dismissed. If it is not, the propositional

content of the sentence is checked for c o n s i s t e n c y relative to

the knowledge base. If it is consistent, the knowledge base is

updated with the new information. If it is not, an assessment is

made as to whether a revision of existing knowledge is called for

in the light of the new information, or whether the new inform

ation is 'wrong', given the validity of the knowledge base. In

the former case, the entire knowledge base is revised, including

the set of possible Inferences that can be drawn from it. In the

latter case, the new information is again rejected (or disputed).

This, in barest outline, is the mechanism that Johnson-Laird

(1983:Ch.l5) takes as the foundation of his theory of how we cre

ate 'mental models' of our environment.

- 138 -

3.4. The r e l a t i o n s h i p b etw een language and r e a l i t y

If the representation hypothesis as formulated in (2) is convinc

ing, then its inherent thesis of the relationship between lang

uage and world must be too. Winograd & Flores describes the lat

ter thus:


(11) The rationalistic tradition regards language as a system of

symbols that are composed into patterns that stand for

things in the world. Sentences can represent the world truly

or falsely, coherently or incoherently, but their ultimate

grounding is in their co r re sp o n d e n c e with the states of af

fairs they represent. This concept of correspondence can be

summarized as:

1. Sentences say things about the world, and can be either

true or false;

2. What a sentence says about the world is a function of the

words it contains and the structures into which these are

combined;

3. The content words of a sentence (such as its nouns,

verbs, and adjectives) can be taken as denoting (in the

world) objects, properties, relationships, or sets of

these.

- 139 -

This is a somewhat simplified account, which I will attempt to

show by means of a slightly expanded paraphrase. Sentences are

said to r e p r e s e n t 'states of affairs', because they say something

about how the world is organized. If the state of affairs that a

sentence represents can be found in the world, then the sentence

is true, otherwise false. A state of affairs is a delimited col

lection of objects which have particular properties and between

which particular relations hold. A sentence represents a state of

affairs just in case the words or phrases of the sentence, and

the structure of which they are a part, refer to those objects


that make up the state of affairs, and specify their properties

and mutual relations.

No wonder that Winograd 6 Flores balk at this. The paraphrase im

plies that there is permanence to a 'true' way in which a sen

tence represents a state of affairs, that a state of affairs can

be identified as the one 'missed' by a particular sentence, that

there is an a priori global but finite set of objects, properties

and relations which we all share a common knowledge of, and which

we all have equal (great or small) possibilities of describing

'correctly'. O'Connor (1975) provides further evidence of the

insufficiency of the above account of the correspondence theory

of truth.

However, a thesis of a c o r re sp o n d e n c e relation between language

and world need not be tied to such a restrictive formulation.

There is nothing in a correspondence thesis that prevents ref

erence to the utterance situation as the constitutive element of

linguistic communication. There is nothing to prevent a general

account that Incorporates speech act theory and correspondence

theory. And there is absolutely nothing to prevent us from re

jecting the simplistic idea of dependency between states of af

fairs and linguistic expressions that forms part of the above

characterization. This leads to the next digression, on universes

of discourse.

- 140 -

3 . 4 . 1 . D i g r e s s i o n : The u n iv e r s e o f d is c o u r s e

Proposition 5.6. in Wittgenstein's T r a c ta tu s states: "D ie Grenzen

m ein er Sprache bedeuten die Grenzen meiner Welt". A possible in


terpretation of this proposition is that language is what con

stitutes our world. There are other possible interpretations. I

shall provide one more below.

The current interpretation leads to the assumption that the world

which language immediately constitutes, is not the actual, phys

ical world, but rather an abstract, a model, or - as we shall

call it - a u n iv e r s e o f d is c o u r s e or, equivalently, a d is c o u r s e

u n i v e r s e .

Discourse universes are private universes, but we can share one

or more of them, in part. Long married couples - who are often

held to be capable of wordless communication - will, in this jar

gon, share a large portion of their overall universe of dis

course. Universes of discourse may be large and small, and may be

more or less densely populated. We can operate on the basis of

more than one universe of discourse at the same time, and we can

shift from one to another with perfect ease between conversations

and even during conversations. And lastly, universes of discourse

are intentional, in Searle's (1983:1) sense of *intentionality':

Intentionality is that property of many mental states

and events by which they are directed at or about or of

objects and states of affairs in the world.

- 141 -

The main idea is that language use implies continuous creation,

revision, and deletion of universes of discourse for the parti

cipants of the conversation. The overall linguistic mechanisms

for these purposes are what I have previously styled 'the refer

ential properties of language' (Thrane 1980). Revision and updat

ing of the current discourse universe is in the main associated


with Identitive, generic and qualitative features (determining

definiteness, specificness, certain aspects of conditionality,

and referent typology), whereas shifts between universes of

discourse are typically associated with presentative and parti

tive features (determining discourse 'scope', various aspects of

definite and indefinite quantification, and other aspects of con

ditionality). The conditionality of referential expressions is

the general mechanism for establishing the 'laws' that hold in a

universe of discourse.

On this view, an utterance becomes a symptom of the state of the

speaker's current universe of discourse, where 'symptom' is in

tended in the technical sense of Lyons (1977:108) as ' a sign or

signal which indicates to the receiver that the sender is in a

particular state'. It is incumbent on the receiver, on the basis

of his interpretation of the symptom, to gain insight in the

state, perhaps to adapt his own universe of discourse accord

ingly, or to try to persuade the speaker to revise his. Mutual

understanding can, under the same view, be regarded as a progres-

sional striving towards the greatest possible congruence between

the current discourse universes of speaker and hearer, through

cooperation and negotiation, for example about the proper defini

tion or Interpretation of a word, or determination of the refer

ence of an expression.

- 142 -

The correspondence relation which is being championed here is a

function from a universe of discours to a state of affairs. And

accepted truths are those special cases in which the universes of

discourse of many (and hopefully, of all) map into the same state

of affairs. This does not prevent 'truth' from being the property


of just one person - in the sense that the relevant state of his

universe of discourse may map into a state of affairs that, over

time, will be mapped by the universe of discourse of society at

large. This is the only kind of ’absolute truth' that the views

taken here will sanction.

3.5. The c o r r e l a t i o n b etw een e x p r e s s io n and c o n te n t

The last step in this account of the various stages of the rep

resentation hypothesis is the question of the connection between

expression and content, or meaning. The two currently most fa

voured bids as to what meaning is, derive from model-theoretic

semantics and Situation semantics. Both attempt to characterize

meaning in a way that enables them to explain the relation be

tween language and reality, both set off from Frege's distinction

between sense (intension) and reference (extension), and both a-

vail themselves of a logical formalism to represent meaning. But

from this point on they part company.

- 143 -

Model-theoretic semantics has taken over Leibniz' notion of 'pos

sible worlds', and defines truth relative to that. The intension

of a sentence is a function from possible worlds to truth-values,

whereas the intension of terms and predicates is a function from

possible worlds to, respectively, objects and sets. The extension

of a sentence is a truth-value, whereas the extension of terms

and predicates is, respectively, objects and sets. So the meaning

(intension) of a linguistic expression is those properties of the

expression which in any conceivable situation determine what

language external objects or sets the expression refers to in

that situation, or - if the expression is a sentence - if the


expression is true in the situation. More briefly: the intension

of a sentence is the set of conditions that has to be satisfied

in some possible world for the sentence to be true in that world.

Model-theretic semantics is fundamentally intensional, and it

represents intensions by means of predicate calculus formulae.

In Situation semantics (Barwise 6 Perry 1983), meaning is a rela

tion between types of situations. This view stems from the recog

nition that although there is a principled difference between

'natural' carriers of meaning ('signs'), and 'unnatural' or 'con

ventional' carriers of meaning ('symbols') - illustrated by the

difference between 'smoke means that something is burning' and

'"something is burning" means that something is burning' - then

both instances involve the transmission of information. Utterance

situations, in other words, belong to a type of situation in

which the transmission of information is based on symbols and

their interpretation. Situation semantics is fundamentally ex-

tensional, and its representation of meaning is in fact a rep

resentation of situation types, in a formal set theoretic nota

tion.

- 144 -

Even though both model-theoretic and Situation semantics seek to

explain the relationship between language and 'the world' through

the development of formal notations for the meaning of natural

language, both theories suddenly lose sight of language. The

model-theoretic solution to the overall problem has the con

sequence that meaning is divorced from language. If intension is

a function from possible worlds (or, in more recent developments,

states of affairs indexed for time) to extensions, then the ling

uistic expression has disappeared and must be reintroduced by a


general interpretation function from expressions to intensions.

The Situation semantic solution fails to draw what to me appears

to be a basic typological distinction between utterance situa

tions (situations created by the making of an utterance), and

other situations (typically situations d e s c r ib e d by making an

utterance). The distinction is based on the notion of intention-

ality. Utterance situations occur whenever utterances are made;

but utterances are the outward manifestation of intentional

states of various sorts. Described situations, on the other hand,

are not intentional. They are rather ' extentional' in the sense

of being the goals at which some intentional states are directed.

- 145 -

Quite apart from such theory-dependent problems, the two semantic

theories share a common problem with any other theory that sub

sumes attempts to represent natural language meaning by means of

a formal notation. The creation of any formal representation of

the meaning of a natural language sentence amounts to nothing but

a claim of synonymy between two material expressions, of course.

The problem that must be faced by all semantic theories that rely

on formal representations of meaning, is that of interpretabil

ity. Even if the formal representation meets all requirements of

syntactic well-formedness. Internal consistency, etc., it is, in

the last resort and in principle, only Interpretable through the

natural language sentence with which it is claimed to be synon

ymous. The problem of interpretability stems from the non-sym-

bollzability of symbols, qua symbols, commented on above (2.1). I

take this to be the onus of the second Interpretation of Wittgen

stein's famous proposition, promised above: my world is only ac

c e s s i b l e through my language. The consequence of this interpreta

tion is that natural language is in fact the only efficient mean

ing representation schema!


3 . 5 . 1 . D ig r e s s io n : C om putation al meaning r e p r e s e n ta t io n

Does this conclusion mean that the representation hypothesis has

foundered as a serious explanatary basis for communication,

reasoning, and knowledge organization? I don't think so. For even

if two well-merited and well-developed semantic theories face

problems with respect to their capacity for giving a global char

acterization of the meaning of natural language sentences, both

have developed techniques and insights that can be brought to

bear on concrete projects that aim at exploring man-machine com

munication in natural language. This has never been the primary

motivation for model-theoretic semantics nor for Situation sem

antics.

Such a narrower approach to the problem will have to set off from

a set of assumptions specifically geared to, and delimiting, its

scope.

First of all, a distinction parallel to Searle's distinction be

tween 'weak' and 'strong' artificial intelligence is called for

within the computational study of language. 'Weak', or 'objec

tive' computational linguistics is characterized by

- regarding language as an object of study;

- regarding the computer as a tool;

- regarding a program as a hypothesis of language structure.

- 146 -

'Strong', or 'subjective', computational linguistics, in con

trast, is characterized by


- regarding language as a medium of communication;

- regarding the computer as a partner in communication;

- regarding a program as a hypothesis of communicative

interaction.

The ultimate goal of the project, on this distinction, falls

within 'strong' computational linguistics - and it is still a

moot point, to what detailed extent, and to what heuristic level,

'weak' computational methods should enter into it (nature of the

parsing required, level of morphological refinement, etc.).

- 147 -

Secondly, it is a limitation that computers, on the 'strong' view

above, meaningfully enter into utterance situations only, indeed,

utterance situations of an impoverished kind: there can (so far)

be no reliance on suprasegmentals, no reliance on gestural or fa

cial information, there is a limited range of language functions,

etc. Thus, 'addressing' a computer is, invariably, an attempt to

gain access to its universe of discourse, either with a view to

changing it or to get documentation of its current state in the

form of an appropriate responsive action. If this action is to be

based on the computer's 'understanding' of the meaning of a nat

ural language input, meaning in this context is a function from

an utterance situation into a discourse universe. This assumption

trades on a combination of the functional and relational views of

meaning characterizing, respectively, model-theoretic and Situ

ation semantics. It further enhances referential and conative

meaning, but deliberately leaves out of account aspects of emo

tive, phatic, metalingual, and poetic meaning.


Thirdly, the process of 'understanding' is further segmented into

a process of 'deciphering' and one of 'interpreting', in the fol

lowing way: 'deciphering' is a function from an u tte r a n c e to a

universe of discourse on the grounds of the 'code', whereas 'in

terpretation' is a function from one universe of discourse to

another. 'Decipherment' will yield a rough approximation to what

the hearer takes as the speaker's current universe of discourse,

'interpretation' will work on this to yield a universe of dis

course more finegrained, and reflecting the hearer's conception

of what the speaker 'has in mind'. Matters of poorly understood

words or phrases will be dealt with by 'decipherment', matters of

failing reference, inconsistency, etc. by 'interpretation'. 'In

terpretation' in this framework is closely akin to the computa

tional view of it (above, 3.1.), in that it involves one or more

processes to be executed by the content of a universe of dis

course .

- 148 -

Finally, as already mentioned, extension is regarded as a func

tion from discourse universes to d e s c r ib e d situations. Whether

the computer in this connection can be said to possess a 'true'

image of the world is a question which is in principle no differ

ent from the question whether we can; it depends on whether our

universe of discourse maps into a factual situation. And this in

turn depends on the nature, quantity and quality of the knowledge

we had access to during its establishment.


- 149 -

4. Final remarks

It has not been my primary concern in this paper to try to fals

ify the representation hypothesis, which I personally find at

tractive. It has been my concern, on the other hand, to present a

series of aspects, interpretations, and consequences of it which

in due course may make it falsifiable. For if it turns out that

the more restrictive semantic program outlined throughout the

last three digressions can be carried through, w ith o u t any indi

cation that the same semantic processes characterize interperson

al communication, then there is reason to believe that the re

presentation hypothesis as formulated in (2) is either too strict

or wrong. However, one of the results may well be acceptance of

the view that the 'nature' of symbols is to be found among the

class of topics of which Wittgenstein said:

Wovon man n lc h t sp rech en kann, dariiber muss man sch w eig en .

One of the fascinations about natural language, however, is that

it is extremely difficult not to use it, even for discussion of

topics that can’t be discussed.


REFERENCES

- 150 -

Barwise, J. & Perry, J. 1983. S itu a t io n s and A t t i t u d e s . MIT:

Cambr. Mass. 2nd Printing 1984.

Brachman, R.J. & Levesque, H.J. (eds.) 1985. R eadings in Know

le d g e R e p r e s e n ta t io n . Morgan Kaufmann: Los Altos.

CP = The C o l l e c t e d P a pers o f C h a rles Sanders P e ir c e . Edited by

Charles Hartshorne and Paul Weiss. Cambr. Mass. 1935-66.

(References in text are to Volume and Paragraph)

Deuchar, M. 1984. B r i t i s h S ign Language. Routledge 6 Kegan Paul:

London.

Goodman, N. 1975. Languages o f A r t . MIT: Cambr. Mass.

Haugeland, J. 1981. Semantic Engines: An Introduction to Mind

Design. In Haugeland, J. (ed.) Mind D e sig n . MIT: Cambr. Mass.

3rd Printing 1985, pp. 1-34.

Hobbes. T. 1651. L ev ia th a n . Pelican Books: Harmondsworth 1963.

Hofstadter, D.R. 1982. Metafont, Metamathematics, and Metaphys

ics: Comments on Donald Knuth's Article "The Concept of a

Meta-Font". In Hofstadter, D.R. M etam agical Themas: Q u esting

f o r th e E ss e n c e o f Mind and P a tte r n . Bantham Books 1986, pp.

260-287.


Hookway, C. 1985. P e i r c e . Routledge & Kegan Paul: London.

Johnson-Lalrd, P. 1983. Mental M o d e ls . Harvard UP: Boston.

Knuth, D. 1982. The Concept of a Meta-Font. V i s i b l e Language 16,

3-27.

Levy, D.M., Brotsky, D.C. & Olson, K.R. 1987. Formalizing the

Figural: Aspects of a Foundation for Document Manipulation.

(Draft). Intelligent Systems Laboratory, Xerox Palo Alto

Research Centre. 3333 Coyote Hill Road. Palo Alto, Ca. 94304.

Lyons, J. 1977. Semantics. Cambridge UP: Cambridge.

Newell, A. 1980. Physical Symbol Systems. C o g n it iv e S c ie n c e 4,

135-183.

Newell, A. 1982. The Knowledge Level. A r t i f i c i a l I n t e l l i g e n c e 18,

87-127.

Newell, A. & Simon, H.A. 1976. Computer Science as Empirical

Inquiry: Symbols and Search. In Haugeland, J. (ed.) Mind

D e sig n . MIT: Cambr. Mass., 3rd Printing 1985, pp. 35-66.

O'Connor, D.J. 1975. The c o r re sp o n d e n c e th e o r y o f tr u th . Hut

chinson: London.

- 151 -

Pylyshyn, Z.W. 1984. Com putation and C o g n it io n : Toward a

Foundation f o r C o g n it iv e S c ie n c e . MIT: Cambr. Mass., 2nd

Edition 1985.


Searle, J. 1980. Minds, Brains, and Programs. In Haugeland, J.

(ed.) Mind D e sig n . MIT: Cambr. Mass., 3rd Printing 1985, pp.

35-66.

Searle, J. 1983. I n t e n t l o n a l i t y . Cambridge UP: Cambridge, 5th

Printing 1987.

Southall, R. 1987. A basis for the description of machine-written

documents. (Draft). Intelligent Systems Laboratory, Xerox Palo

Alto Research Centre. 3333 Coyote Hill Road. Palo Alto, Ca.

94304.

Thrane, Torben 1980. Referential-semantic a n a l y s i s : A s p e c ts o f a

th e o r y o f l i n g u i s t i c r e f e r e n c e . Cambridge UP: Cambridge.

Waterman, D.T. 1986. A Guide to E x p e rt S y s te m s . Addison-Wesley:

Reading, Ma.

Winograd, T. & Flores, F. 1986. U n d ersta n d in g Com puters and Cog

n i t i o n . A New F ou n d a tion f o r D e sig n . Ablex: Norwood, N.J.

- 152 -

Wittgenstein, Ludwig 1921. T r a c ta tu s L o g ic o -P h ilo s o p h ic u s . Rout-

ledge & Kegan Paul: London, 1966.


- 153 -

THINK tab: forehead dez: index finger extended

from closed fist sig: contact with tab

KNOW tab: forehead dez: thumb extended from

closed fist5ig: contact with tab

CLEVER tab: foreheaddez: thumb from closed fist sig: movement from right to

left in contact with tab

The structure o f BSL signs


- 1 54 -

(--------- ^■il ■ j

i 'V>r̂ rL S I

7^ .̂ ^^ •y<Lon:5D ^ii d d r fc r ijft /^9c.


abstract - association for computational linguisticsmiller (1984) called 'informavores',...

Documents